Giraphx: Parallel Yet Serializable Large-Scale Graph Processing

Abstract

Abstract. Bulk Synchronous Parallelism (BSP) provides a good model for parallel processing of many large-scale graph applications, however it is unsuitable/inefficient for graph applications that require coordination, such as graph-coloring, subcoloring, and clustering. To address this problem, we present an efficient modification to the BSP model to implement serializability (sequential consistency) without reducing the highlyparallel nature of BSP. Our modification bypasses the message queues in BSP and reads directly from the worker's memory for the internal vertex executions. To ensure serializability, coordination is performedimplemented via dining philosophers or token ring-only for border vertices partitioned across workers. We implement our modifications to BSP on Giraph, an open-source clone of Google's Pregel. We show through a graph-coloring application that our modified framework, Giraphx, provides much better performance than implementing the application using dining-philosophers over Giraph. In fact, Giraphx outperforms Giraph even for embarrassingly parallel applications that do not require coordination, e.g., PageRank.

High-Level Programming Abstractions for Distributed Graph Processing

“…Year Programming Model Execution Communication Pegasus [62] 2009 MapReduce S Dataflow Pregel [2] 2010 Vertex-Centric S Message-Passing Signal/Collect [10] 2010 Scatter-Gather A, S Message-Passing HaLoop [69] 2010 MapReduce S Dataflow Twister [70] 2010 MapReduce S Dataflow Piccolo [72] 2010 Partitioned Tables S Shared global state Apache Giraph [36] 2011 Vertex-Centric S Message-Passing Comb. BLAS [31] 2011 Linear Algebra S Message-Passing Apache Hama [53] 2012 Vertex-Centric S Message-Passing GraphLab [1] 2012 Vertex-Centric A, S Shared Memory PowerGraph [7] 2012 GAS A, S Shared Memory Giraph++ [11] 2013 Subgraph-Centric H Message-Passing Naiad [60] 2013 Dataflow A, S, I Dataflow GPS [21] 2013 Vertex-Centric S Message-Passing Mizan [6] 2013 Vertex-Centric S Message-Passing Presto [73] 2013 Linear Algebra S Dataflow Giraphx [64] 2013 Vertex-Centric A Shared Memory X-Pregel [66] 2013 Vertex-Centric S Message-Passing LFGraph [32] 2013 Vertex-Centric S Shared Memory SociaLite [29] 2013 Datalog Extensions S Message-Passing Trinity [28] 2013 Vertex-Centric A, S Message-Passing, Shared Memory Graphx [25] 2014 GAS S Dataflow GoFFish [12] 2014 Subgraph-Centric H Message-Passing Blogel [13] 2014 Vertex-Centric, Subgraph-Centric H…”

Large-Scale Graph Processing Using Apache Giraph

“…Since Giraph's computing model is directed and distributed; it does not naturally support undirected graphs or have the ability to synchronize value assignment to adjacent vertices. To solve this problem, researchers tend to modify Pregel API to serialize the value assignment process through the master compute, as suggested by the work in [48,49]. In this section, we show how to handle undirected graphs and discuss non-traditional fine grain vertex synchronization without the need to modify Giraph API.…”

Towards Scale-out Capability on Social Graphs

“…It has been noted that the vertex-centric BSP model is unsuitable for the graph applications that require coordination, e.g., graph coloring. Lately, Giraphx [30] presented a modification to the BSP model to meet the aforementioned requirement by categorizing vertices as border and internal vertices. The modification allows the direct read to worker's memory for internal vertices.…”

“…Unfortunately, MapReduce, the de-facto big data processing model, and its associated technologies such as Pig [19] and Hive [31] are often ill-suited for iterative computing problems on large graphs. As a result, several graph-parallel projects [1,17,18,22,24,27,30,32,36], have been proposed to facilitate data analysis on large graphs. These projects can be generally classified into two categories:…”

“…The function vertex.compute() receives messages via the incoming edges from the previous superstep, processes the messages such as an aggregation, and sends messages via the outgoing edges for the next super-step. Pregel inspires several research and open-source projects [1,22,24,27,30,32,36].…”